ByteDance's SeedRealtime Can Watch, Listen and Speak at Once: If You're in Doubao

ByteDance's Seed research organization launched SeedRealtime on August 5, describing it as a "native audio-visual full-duplex" large language model that fuses sight, sound and text in a single architecture and holds a live conversation over continuous multimodal streams. The model is already running inside Doubao, ByteDance's consumer assistant app, where it is free to use on iOS and Android.

Related reading: Google's Gemini 3.8 Live Reasons Out Loud While It Talks · Anthropic's Claude Haiku 5.5 Is Its Cheapest Small Model Yet.

The phrase "full-duplex" is doing a lot of work. In a normal voice assistant, you speak, the system detects that you have stopped, then it answers. That turn-taking is orchestrated by an external component (a voice-activity detector) bolted onto a pipeline of separate speech recognition, language and speech-synthesis modules. Full-duplex means the model listens and speaks at the same time, tracking who is talking and when to interject the way a person does. SeedRealtime's claim to novelty is that it does this, and does it with video as well as audio, inside one end-to-end model.

Three tiers of "real-time"

It helps to separate the field, because the marketing blurs it. One useful framing, offered by the analyst site ORCARouter, sorts realtime assistants into three tiers. The first is half-duplex voice: most assistants, including Google's Gemini Live, which can take camera and screen input but still waits its turn. The second is audio-only full-duplex, where the system listens and speaks simultaneously but has no eyes, ByteDance's own Seeduplex, OpenAI's GPT-Live and xAI's Grok Voice fall here. The third is audio-visual full-duplex, where continuous perception and expression span both modalities. SeedRealtime is ByteDance's bid for that third tier, and the company says it is the first at this level to be deployed at large scale.

Whether that last claim holds will take time to assess, because there is nothing public to compare it against. But the technical description is concrete. SeedRealtime uses a unified audio-visual modeling framework in which perception, understanding, decision-making and response generation all happen in one model, rather than being handed between specialist modules. ByteDance argues this cuts the information loss and error accumulation that come from a cascade, where each handoff can drop timing cues, tone, or context.

The practical consequence is that the model uses the live scene to resolve ambiguity. It can disambiguate a homophone by reading what is on screen, and it can react proactively rather than waiting to be addressed. In demos, ByteDance shows the model flagging a museum exhibit a user had mentioned the moment it comes back into the camera's view, and catching a mistake during a live espresso pour. It also manages its own turn-taking by tracking the scene, the speaker, pauses and background chatter, rather than relying on an external trigger.

A milestone you can't build on

The most important fact about SeedRealtime for developers is what it lacks. As of launch, there is no public API, no model ID, no pricing and no developer documentation. Reports noted that the model is reachable only through the Doubao app, and that ByteDance's commercial API arms, Volcano Engine in China and BytePlus internationally, do not list it. The company says the model is "fully rolled out," but that refers to deployment inside its own products, not to an endpoint anyone can call.

That is a familiar pattern from ByteDance, and a recurring one in the realtime-AI field. Frontier multimodal systems tend to reach consumers first, inside a flagship app, and reach builders later, if at all. ORCARouter noted that ByteDance shipped the audio-only Seeduplex into Doubao in April 2026, roughly three months before OpenAI's GPT-Live reached ChatGPT's voice mode. If that precedent holds, a SeedRealtime API is a matter of when, not whether, but the timeline is ByteDance's to set.

The gap between "fully rolled out" and "available to developers" is the gap between a demo and a platform.

For teams trying to plan around realtime interaction, this is the crux. A model that exists only inside a consumer app cannot be evaluated for latency, cost, refusal behavior or reliability under production load. The competitive comparison that matters (SeedRealtime versus a callable audio full-duplex model) cannot yet be run. What builders can do today is prototype against the APIs that do exist and treat SeedRealtime as a directional signal about where the category is heading.

Privacy and availability caveats

Two caveats deserve more attention than launch coverage gave them. The first is geography and compliance. SeedRealtime is a feature of a mainstream Chinese consumer app, reached through Doubao's combined camera-and-microphone interface, and it operates under Doubao's existing moderation and legal rules rather than a model-specific policy. That shapes what it will and will not discuss, and it means availability outside China is not straightforward.

The second is data handling. A session that continuously ingests camera and microphone input is a different privacy proposition from a text chatbot. ByteDance's general Seed privacy policy, per one technical summary, says the company collects content a user uploads plus attached file metadata, but there is no SeedRealtime-specific retention window published, and no system card or red-team summary. For a model that watches your room and listens to your conversation, that is a gap worth naming, even if it is a common one across the industry.

Why end-to-end matters, technically

Cascaded systems are easier to build but harder to make feel natural. In a pipeline, an automatic speech recognizer turns audio into text, a language model decides what to say, and a text-to-speech model says it. Each stage introduces delay, and each stage can lose information the next one needs, prosody, hesitation, a glance at an object. The turn detector sits outside the model entirely, which is why so many assistants either interrupt too eagerly or leave awkward silences.

An end-to-end, full-duplex model folds all of that into one network trained on the joint task. In principle, it learns the timing cues a cascade discards, and it can generate backchannels (the "mm-hm" and "right" that humans sprinkle into conversation) without a separate rule engine. The reason this is hard is data. Synchronized audio-visual-conversational data at scale is scarce and expensive to label, and there is no standard benchmark that says whether a model has the rhythm right. That vacuum is exactly why ByteDance's claim is hard to verify: the field has not agreed on how to measure the thing SeedRealtime is selling.

Serving is the second hard part. Continuous listening and speaking is a streaming workload, not a request-response one, and the compute has to stay resident for the length of the conversation. That makes full-duplex interaction markedly more expensive per user than a text chatbot, which is one reason it tends to ship inside a large consumer app that can absorb the cost, a strategy ByteDance has followed before.

A consumer-first playbook

ByteDance's decision to keep SeedRealtime inside Doubao is consistent with how it handled earlier Seed models. Doubao gives the company a distribution channel measured in the hundreds of millions, and it lets ByteDance gather the messy real-world conversations that improve a realtime model without exposing the system to third-party integration risk. If the broader ambition is a family of personal AI devices, as the company's hardware and wearables work suggests, then owning the full-duplex conversational layer is a prerequisite rather than a side project.

The cost of that approach is ecosystem gravity. When the strongest realtime models are reachable only inside one vendor's app, builders route around them, toward providers that offer a callable endpoint. ByteDance has traded near-term developer mindshare for control of the experience. Whether that trade looks smart a year from now depends on whether the API arrives before the category standardizes without it.

Why this matters

SeedRealtime points at where conversational AI is going next. The last two years of voice assistants were about making turn-taking feel less robotic. The next phase is about removing the turn-taking machinery altogether, so that a model simply participates in a scene. Doing that with video, not only audio, is a meaningful step: it makes the assistant aware of the physical context around a conversation and capable of speaking first when something visible warrants it.

The strategic stakes are high for ByteDance. Full-duplex interaction is expensive to serve and hard to do well, and owning the model, the app and the user relationship gives the company an advantage in tuning the experience. But the same closed posture that lets it move fast also keeps it out of the developer ecosystem where agent infrastructure is being standardized. If SeedRealtime's approach is right, the API release will be the event that matters.

None of this makes SeedRealtime unimportant. It is significant precisely because it collapses a boundary the industry treated as fixed for years: that a machine must wait its turn. Watching and speaking at the same time is a small thing in a demo and a large thing in daily use. It is the difference between an assistant you summon and one that is simply present. That is closer to a product category than a feature, and it explains why ByteDance prioritized shipping it over monetizing it.

What to watch

Watch for a SeedRealtime model ID on Volcano Engine or BytePlus, and watch for the latency number that comes with it, end-to-end latency is the metric that separates a gimmick from a product in realtime interaction. Watch, too, for Microsoft's MAI realtime speech model and Google's next Gemini Live update, both of which are circling the same territory. And watch Doubao retention: the honest test of full-duplex video interaction is whether users keep the camera session open once the novelty wears off.